Jerusalem, 14 October, 2026 (TPS-IL) — Israeli researchers have developed an artificial intelligence model that reconstructs images from brain activity using about one hour of brain scans from a new participant, while identifying previously unknown functional divisions in brain regions involved in visual processing.
The model, called Brain-IT, was developed by a team led by Prof. Michal Irani at the Weizmann Institute of Science. Unlike existing systems that require dozens of hours of brain scans to adapt to a new person, Brain-IT identifies patterns shared across different people, allowing it to reconstruct images with greater accuracy and less training.
Beyond reconstructing images, the technology could help scientists understand how the brain organizes visual information. It could also advance the development of brain-computer interfaces that require less training, though practical medical applications are still a long way off.
“There exist nowadays models that translate brain activity into images, and they can even produce impressive reconstructions that preserve the semantic meaning of the image reasonably well,” Irani said. “However, they tend to make mistakes in basic features such as composition and color.”
“The new model we developed outperforms them in reconstructing both the content of the image and its details,” she added. “What’s more, while every other model requires dozens of hours of brain scans to learn to ‘read’ a new person, our model needs only one hour.”
The research was selected for presentation this year at the International Conference on Learning Representations, a leading artificial intelligence conference.
Teaching AI to Read the Brain
A major obstacle in this field is the shortage of training data. Researchers need images paired with functional magnetic resonance imaging (fMRI) scans, which measure changes in brain activity as people view pictures. Collecting such data requires participants to spend many hours in MRI scanners.
The publicly available Natural Scenes Dataset used in the study contains about 73,000 image-scan pairs from eight participants, each of whom viewed thousands of images over 30 to 40 scanning sessions.
To overcome this limitation, the team developed two complementary systems: an encoder that predicts brain activity from an image and a decoder that reconstructs an image from brain activity.
The encoder generated predicted brain scans for images that participants had never viewed inside an MRI scanner. The decoder then tried to reconstruct those images from the predictions, allowing the systems to learn from additional examples without requiring new scans for every image.
The researchers divided the brain into roughly 40,000 tiny three-dimensional units called voxels and compared their responses to different images across individuals. This revealed recurring patterns linked to visual features such as color, location and recognizable objects.
The encoder identified 128 functional regions that were shared across people and appeared to play specific roles in visual processing. Some were already known to scientists, while others were new.
One finding involved the parahippocampal place area, a brain region involved in processing places and scenes. The researchers found evidence of a division of labor within the region, with one part responding to indoor scenes and another to outdoor scenes.
“The encoder’s strength lies in its ability to identify regions that perform similar functions in different brains, even when their anatomical locations differ,” Irani said.
In the near term, Brain-IT could help researchers study visual processing while reducing the need for costly, time-consuming scans. Its ability to identify patterns shared across individuals could also help scientists develop brain-computer interfaces that require less training to adapt to new users. Further research is needed